Agent for OS

Kleinia brings the adaptability of agents to resource scheduling. It understands not just the workload but what you and the system need, and switches scheduling policies as needs change.

Adaptive scheduling

The resource agent reads the character of each workload, picks the best CPU, I/O and page cache policies for it, and switches them in real time as the load changes.

Policies that follow the load: LAPS recognizes the current workload in real time and switches between scheduling policies. In the paper’s experiments it improves performance by 8.91% on average over EEVDF, the Linux default, and cuts response latency by up to 91.7% in latency-sensitive tasks. The code is open: lvmiao233/LAPS.
Intelligence decoupled from real time: agents recognize the workload and decide when to switch; only deterministic policies run in the kernel, so performance never costs safety or overhead. No language model sits on the scheduling path.
Resources in concert: CPU, I/O and page cache follow one policy.

agenttop shows the policies currently in use.

An extensible scheduling substrate

Scheduling policies run in the kernel as pluggable modules: sched_ext for the CPU and cache_ext for the page cache. A compiled policy passes the kernel’s safety check before it loads; policies switch at runtime without rebuilding the kernel, and a misbehaving policy falls back to the default.

Self-evolving

Faced with a workload it has not seen, the system analyses it in the background and updates its scheduling model from measurements on this machine. The foreground is never disturbed.

Shared experience

Scheduling experience gathered on one device can be shared with other devices in the community, so new machines adapt faster.